Reconstruction and Upscaling of 3D Models from Single or Multiple Views

Abstract

In recent years research related to 3D reconstruction from 2D images has gained traction and several approaches have been introduced. However, most conventional methods for 3D reconstruction are time consuming and tedious. Additionally, they produce low resolution results and have their own limitations. Our approach attempts to resolve these limitations by using a modified encoder-decoder architecture which generates a low resolution 3D coarse volume from a set of 2D images of an object. In order to improve the quality of the generated model, a pseudo high resolution 3D volume is generated by upsampling the low resolution volume which has multiple missing features. Parallelly, RGB-D images from different angles are generated using the Blender software. Furthermore, these RGB-D images are upscaled to high resolution images using a CNN-image upscaler and a depth map is extracted. These newly generated depth values assist in identifying the missing features from the pseudo 3D volume thereby generating a final high quality 3D coarse volume. Our results show that this approach outperforms the existing methods.